Closed hoeflerb closed 8 months ago
Just to clarify, the principal aim here is to do a nested model comparison. For example, to compare two models with and without an interaction term included.
Thanks for the inquiry, @hoeflerb!
We recently reimplemented scCODA as part of the pertpy package with improved functionality and API. This will also be the maintained version of scCODA in the future.
There, you can also create an arviz object from the inference result, which actually includes the log-likelihood. I just tried calculating the WAIC on our tutorial dataset, and it seems to work:
Hi,
From the advanced tutorial documentation, scCODA appears to use the arviz InferenceData object to store the fitted model. This is handy and convenient because arviz provides a lot of useful utilities for generating model summary stats, diagnostic plots, etc. Unfortunately, the scCODA model output appears to not include the log likelihoods, so functions like arviz.waic() and arviz.loo() don't work.
Is there a way to include these in the model output (maybe by a parameter to sample_hmc()) or to compute them independently?
Thanks!